A hidden Markov model-based K-means time series clustering algorithm

Abstract

Aimed at some shortages in the existing time series clustering methods based on hidden Markov model(HMM), such as longer sequence and equal length, a hidden Markov model-based k-means time series clustering algorithm is proposed, whose objective function is the joint likelihood function. At first, an initial partition is obtained by unsupervised clustering of the time series using dynamic time warping (DTW), then HMMs are built from it, and the initial clusters serve as input to a process that trains one HMM on each cluster and iteratively moves time series between clusters based on their likelihoods given the various HMMs.